Code of paper "Dual Principal Component Pursuit for Learning a Union of Hyperplanes: Theory and Algorithms", AISTATS 2021

The code has been tested to run in MATLAB R2018b.

• RSGM_demo.m produces Figure 2 in the paper, which illustrates the linear convergence of the Projected Riemannian Subgradient Method with different geometrically diminishing factors.
• compare_KSS.m produces Figure 3 in the paper, which compares DPCP-KSS, CoP-KSS and PCA-KSS per iteration in terms of their clustering accuracies (same initialization).
• run_all_example.m provides a quick example of running all methods once
  • Settings:
    • ambient dimension D=4
    • number of hyperplanes K=2
    • number of inlier points N1=N2=200
    • ouliter ratio M/(M+N)=0.3
  • It runs the following methods:
    • MKF
    • SCC
    • EnSC
    • SSC-ADMM
    • SSC-OMP
    • DPCP-KSS, CoP-KSS, PCA-KSS
    • DPCP-EKSS, CoP-EKSS, PCA-EKSS
    • DPCP-CoRe-KSS, CoP-CoRe-KSS, PCA-CoRe-KSS

@inproceedings{ding2021dual,
  title={Dual Principal Component Pursuit for Learning a Union of Hyperplanes: Theory and Algorithms},
  author={Ding, Tianyu and Zhu, Zhihui and Tsakiris, Manolis and Vidal, Rene and Robinson, Daniel},
  booktitle={International Conference on Artificial Intelligence and Statistics},
  pages={2944--2952},
  year={2021},
  organization={PMLR}
}